EU Bans Workplace Emotion AI From Feb 2025; HireVue, Microsoft Retreated Earlier, Inno3 Traces Arc

Candidate sits alone, webcam on, algorithm watching: that was HireVue’s video-interview pipeline, which claimed more than 19 million interviews for over 700 clients before the company made it known in January 2021 that it had removed facial analysis from its assessments, and that it had actually done so in March of the year before — and the science behind it has been crumbling ever since.

The retreat is now industry-wide and, as of February 2, 2025, legally mandated in Europe.

The EU’s AI Act bans systems that infer emotions in workplaces and schools, with narrow exceptions for medical or safety reasons, and the reform approved in the summer of 2026 that pushed high-risk compliance obligations to December 2027 left those prohibitions untouched. inno3.it traced the full arc, from a flawed 1967 psychology experiment to a continent-wide ban, and the picture it assembles is brutal for the emotion-AI vendor category.

The telling detail is who blinked first. The companies selling facial emotion recognition pulled their own products before Brussels said a word.

HireVue dropped the face. Then Microsoft followed

HireVue spent years scoring candidates on smiles, furrowed brows and micro-expressions while an algorithm also listened to their words. In January 2021 the company announced it had removed facial analysis from its assessments — and that it had actually done so the previous March. The official rationale, as inno3.it reported, deserves a close read: internal research showed language analysis had become more predictive, and visual analysis no longer added significant value.

Translation: the vendor of face-reading concluded that the face wasn’t the signal.

It wasn’t purely a data-driven epiphany. The CEO admitted to Wired that protests had weighed on the decision, and an algorithmic-fairness consultant had resigned in 2019 over exactly this feature.

Microsoft made it two in June 2022, retiring Azure Face’s capabilities that attributed emotional states to faces, keeping them only in a handful of accessibility tools. Natasha Crampton, chief responsible AI officer, cited the lack of scientific consensus on the very definition of emotion.

AI Act bans workplace emotion tracking

The AI Act’s prohibition, in force since February 2, 2025 under Article 5, covers the entire employment relationship, hiring included. The European Commission’s guidelines, published February 4, 2025 and non-binding, spell out the scenarios: a call center using webcams and voice recognition to track operator anger is banned, and so is monitoring a team’s emotional climate from video-call audio and imagery. Measuring employee motivation or satisfaction does not qualify for the medical exception.

Two carve-outs survive. Customer emotion analysis remains legal. And, curiously, so is the smile itself: a system may count how often a TV host smiles, as long as it doesn’t conclude the host is happy.

The 7-55-38 myth that launched a thousand dashboards

The belief that face and voice outweigh words predates any software. Corporate training has repeated it for decades with numbers attached: words count 7 percent, tone of voice 38 percent, the face 55 percent. Total nonverbal share: 93 percent. If that were true, you could watch a film in an unknown language and lose only a few details.

The numbers trace to Albert Mehrabian, the UCLA psychologist, and Susan Ferris, Licensed Clinical social worker at Ucla Health, and 1967 papers, according to Inno3. One, with Susan Ferris, sat twenty psychology students in front of a black-and-white photo of one woman’s face while hearing the single word “maybe” spoken with liking, neutrality or dislike by a different woman, then asked how much this laboratory-assembled person liked an imaginary interlocutor. The face weighed roughly one and a half times the voice.

The words couldn’t be measured at all — there was one, deliberately neutral. The 7 percent figure came from a second study that year with Morton Wiener, published in the Journal of Personality and Social Psychology, built on nine single words; the full formula appears in the last paragraph, introduced by a cautious “it is suggested that,” and the authors themselves call it an extrapolation.

The paper was about something else entirely: a theory on the origins of schizophrenia.

Mehrabian spent decades trying to push the toothpaste back. In a 1995 letter he called the idea that words carry just 7 percent of any message absurd, and always limited his findings to messages about feelings and attitudes. When David Lapakko, communication lecturer, checked in 2007 how the formula circulated online, only 4 of 79 websites repeating it expressed any doubt.

What the research actually says in 2026

The scientific consensus Microsoft couldn’t find genuinely doesn’t exist. In 2019, Lisa Feldman Barrett, nota psicologa canadese, and four colleagues reviewed more than a thousand published results on facial movement and emotion in Psychological Science in the Public Interest.

A furrowed brow is not anger; a smile is not happiness; both depend on the person, the culture, the situation. Software, they wrote, can detect a facial movement but not its psychological meaning.

But the question isn’t closed, and presenting it as such would be dishonest. In the same journal issue, Alan Cowen, emotion scientist and AI engineer, and colleagues countered that voice alone distinguishes more than twenty emotions, and that face, body and context together say quite a lot.

Their demand, though, aligns with Barrett’s: the link between an expression and what someone feels must be derived from data case by case, never assumed. On that point the two schools agree.

What survives of Mehrabian is the observation he genuinely made: when tone or face contradicts the words, people believe the tone and the face. Anyone who has received a gritted-teeth “everything’s fine” knows this without percentages.

What enterprise buyers should ask vendors

The percentage doesn’t hold, and neither does the idea that a person can be read like a dashboard. Anyone pitching a platform that measures team engagement or candidate enthusiasm should face three questions, per the reporting by inno3.it: validated on which populations, against what outcome, audited by whom.

In 1967 the honest answer would have been: on twenty studentesse, with a “maybe”.

A number doesn’t become true because it sits in a slide deck. Nor because it sits in an algorithm.

Photo of author

Technology Desk

Technology Desk at Archyde.

Quad Leaders Meet in Manila, Announce AI Cooperation; Analysts Call It ‘Photo Diplomacy